If your model of Donald Trump is "cartoonishly evil and incapable of empathy", then yes, of course you need to adjust your model – but that's a bad description of Adolf Hitler, too. He genuinely cared about the welfare of certain people, and opposed smoking because of the harm it caused those people: if you pegged Hitler as generally pro-death, you'd be wrong. But that does not in any way redeem him, and it shouldn't cause you to update your "Hitler wants to kill a whole bunch of people" prediction.

Suppose it's 1940. You know that Hitler ordered Aktion T4, and conclude that Hitler wants to kill people. Then, you learn that he opposes smoking because he doesn't like it killing people. You shouldn't start doubting that he's the sort of guy to sign mass death warrants: you've learned some information about his internal thought processes, but it's not very useful information if you want to predict his future actions.

"Orthogonal" is subjective. All things are interrelated. That does not mean that our descriptions should be highly-sensitive to noise. Update your internal model of his behaviour, by all means, but if you have predictions that don't require that internal model, consider whether or not this evidence should actually affect those predictions.